HSI strategy analysis¶
strategy:¶
Entry conditions
- At pre-set times (9:15am and 13:45pm), we will make a bet.
- If the price of that pre-set time is higher than previous day close, it will make a Long position; while if the price of that pre-set time is lower than previous day close, it will make a Put position.
Exit conditions
- If the position is winning 80 points (84 points to be exact), it will trigger a trailing stop when it drawdowns by 40 points from the max winning position.
- If the position is not winning over 80 points, it will trigger a stop loss exit with a 40 points loss.
Multichart statistics shows that the strategy was rather unstable across the course of time of backtesting. With from 2018 to 2020H1, it is basically flat or even losing, but during 2020H2, the performance of the strategy was stunning.
- Why is that?
- What kind of market situations will it be flat (and losing)?
- What kind of market situations will be winning?
- If there are chances we can have signals telling us which market is favourable to the strategy? And telling us to or not to enter the market?
- Can we decrease the volatility of the strategy, making it more stable across the course of time?
Overview: What's the win rate of the strategy?¶
In [14]:
df_trades['exit_Signal'].value_counts().sum()
Out[14]:
- Over 1000 trades are analysed here.
In [16]:
df_
Out[16]:
- Overall, 64% of the trades had stopped-loss and only 35% of the trades were winning.
Were short entries more profitable than long entries? No.¶
In [19]:
ax = df_.plot.box(figsize=(7,4))
plt.ylabel('% of trades with stopped-loss in a month');
In [21]:
df_e
Out[21]:
- The percentage of stopped-loss trades do not seem to differ much by whether the trades were long or short entries.
In [22]:
stats.ttest_ind(df_.Buy, df_.Sell)
Out[22]:
- From t-test, there is no evidence that the long and short entries differ in % of trades with stopped-loss
Confirm Multichart results: Did the months in 2020-H2 made more profit?¶
In [24]:
df_monthly_profit.plot.bar( figsize=(12,5) );
- YES. 2020 H2 had more profits compared to periods before and after; only 2020-10 had a loss.
- 2021 Jan had the largest gain
- 2019 Jul had the largest loss
Confirm Multichart results: Did 2020-H2 have less losing trades per month compared to other periods?¶
In [30]:
plt.style.use('bmh')
ax = df_.plot.box()
plt.ylabel('% of trades with stopped-loss in a month');
- YES. 2020 H2 had a lower percentage of trades with stopped-loss in a month.
P.S. 2018 H1 only had 1 monthly data point.
In [31]:
df_exit_stat['% Stop Loss'].plot.bar( figsize=(12,5) );
- The particularly low % of trades with stopped loss in 2020-H2 was not due to any outlier in a particular month, although 2020-Aug was lower in % of trades with stopped loss. So the outperformance of 2020-H2 was not a skewed result.
Investigate trades in 2019-July, which had many losses¶
- Trades in 2019-Jul should indicate more strongly as to what caused the poor performance of the strategy.
Note: Entries are shown as triangles; stop-losses are crosses; short-cover and long-sell are pointers; end of day exits are pentagons
In [38]:
for i, trade in df_trades.query(str_201907).iterrows():
plot_trade(trade, df_bar, my_style)
Investigate trades in 2020-Aug, which had a good number of winning trades.¶
- Trades in 2020-Aug should indicate more strongly as to what caused the better performance of the strategy.
In [39]:
for i, trade in df_trades.query(str_202008).iterrows():
plot_trade(trade, df_bar, my_style)
- From eyeballing the charts, seems like the winning trades happen more often when there were pullbacks in price and then the price continues in the direction of interest. Pullbacks seemed to happen more often in 2020-Aug and less in 2019-Jul.
Did winning trades happen more often when there were pullbacks ?¶
2019-Jul¶
In [40]:
ax = df_trades.query(str_201907)[ df_trades.query(str_201907)['pullback_at_entry'].notna() ].boxplot(by='win_loss', column=['pullback_at_entry'] )
ax.set_ylim(0,100);
In [42]:
stats.ttest_ind(s_win, s_loss)
Out[42]:
- 2019-Jul was a month with many loss trades and this month
- And the winning trades tend to have more pullbacks (~28 pts); however, the difference in pullback between win and loss trades is not statistically significant.
In [43]:
df_trades.query(str_201907).groupby('win_loss')['pullback_at_entry'].describe()
Out[43]:
2020-Aug¶
In [44]:
ax = df_trades.query(str_202008)[ df_trades.query(str_202008)['pullback_at_entry'].notna() ].boxplot(by='win_loss', column=['pullback_at_entry'] )
ax.set_ylim(0,100);
In [46]:
stats.ttest_ind(s_win, s_loss)
Out[46]:
- 2020-Aug was a month with many winning trades
- And the winning trades tend to have more pullback (~34 pts); however, the difference in pullback between win and loss trades is not statistically significant.
In [47]:
df_trades.query(str_202008).groupby('win_loss')['pullback_at_entry'].describe()
Out[47]:
In [48]:
ax = df_trades[ df_trades['pullback_at_entry'].notna() ].boxplot(by='win_loss', column=['pullback_at_entry'] )
ax.set_ylim(0,100);
- Indeed the difference in pullback between win and loss trades is not apparent at all when looking at all the trades from 2018 to 2021-H1.
Note: the boxplot is cropped at the top in order to see the median.
In [50]:
stats.ttest_ind(s_win, s_loss)
Out[50]:
Does more volatile market bring more profit ?¶
In [59]:
ax = df_trades.query(str_b4_2020H2).boxplot(by='win_loss', column=['today_range'] )
ax.set_ylim(0,1000);
- From 2018 to 2020-H1, the winning trades had higher volatility range and this result is statistically significant.
In [61]:
stats.ttest_ind(s_win, s_loss)
Out[61]:
In [62]:
ax = df_trades.query(str_2020H2).boxplot(by='win_loss', column=['today_range'] )
ax.set_ylim(0,1000);
In [64]:
stats.ttest_ind(s_win, s_loss)
Out[64]:
- In 2020-H2, the winning trades had slightly higher volatility range; though this result is not statistically strong enough.
In [65]:
ax = df_trades.boxplot(by='win_loss', column=['today_range'] )
ax.set_ylim(0,1000);
In [67]:
stats.ttest_ind(s_win, s_loss)
Out[67]:
- If volatility is high for the trading day then the strategy is more likely to win and this result is highly statistically significant
Should we place a trade in the afternoon if the volatility in the morning was high?¶
In [68]:
ax = df_trades.query('entry_Signal == "LEpm" | entry_Signal == "SEpm"').boxplot(by='win_loss', column=['morning_range'] )
ax.set_ylim(0,1000);
In [69]:
df_trades.query('entry_Signal == "LEpm" | entry_Signal == "SEpm"').groupby('win_loss')['morning_range'].describe()
Out[69]:
- unfortunately, high volatility in the morning session does not lead to better winnings in the afternoon.
In [95]:
stats.ttest_ind(s_win, s_loss)
Out[95]:
Does today's range correlate with previous day's range ?¶
In [72]:
df_trades[['today_range','prev_range']].corr()
Out[72]:
- There's only 38% correlation between today's volatility and previous day's volatility; not much at all.
In [73]:
df_trades[['today_range','prev_range']].plot.scatter(x='prev_range', y='today_range');
In [74]:
df_trades[['today_range','prev_range']].apply(lambda x: np.log(x)).add_prefix('log_').plot.scatter(x='log_prev_range', y='log_today_range');
Does today's win/loss correlate with previous day's range ?¶
In [76]:
plt.style.use('ggplot')
df_.plot.hist(bins=30,alpha=0.5, figsize=(12,5));
In [77]:
ax = df_trades.boxplot(by='win_loss', column=['prev_range'] )
ax.set_ylim(0,1000);
- Unfortunately, previous day's range does not correlate with next day's win loss of the strategy.
Appendix¶
List of Trades with End of Day Exit¶
In [78]:
df_trade[ df_trade['Signal'] == 'End of Day Exit' ]
Out[78]: